Page MenuHomec4science
Files F99976627

ridge_regression.py
No OneTemporary
Actions

File Metadata

Created
Mon, Jan 27, 13:39

ridge_regression.py
View Options

# coding: utf-8
# In[ ]:
import matplotlib
matplotlib.use('Agg') # Must be before importing matplotlib.pyplot or pylab!
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
pd.set_option('display.max_columns', 100)
# In[ ]:
# import scikit learn packages
from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_squared_error
from sklearn.metrics import r2_score
from sklearn.linear_model import RidgeCV
from sklearn.preprocessing import StandardScaler
from sklearn.preprocessing import PolynomialFeatures
from sklearn.pipeline import make_pipeline
# In[ ]:
from helpers import *
# **Settings**
#
# *Boleans*
# - ``seasonwise``: If set to ``True`` perform 4 different regression splitting the dataset season by season. Otherwise perform one single regression.
# - ``feature_selection``: If set to ``True`` filter only season selected by random forest features selection (80% importance, 13 out of 20 features). Otherwise utilise all the 20 features.
# - ``output_y_only``: If set to ``True`` perform regression considering only ``u_y`` as output. Otherwise consider speed in both $x$ and $y$.
#
# *Parameters*
# - ``rnd_state``: Seed state used in the splitting of the dataset. Default in all algorithms is $50$.
# - ``alphas``: Possible choice of regularization parameter optimized by the ridge regression. Default is `` np.logspace(-10,5,200)``.
# - ``deg``: Degree of the polynomial expansion. Default is $3$.
# - ``train_dim``,``test_dim``,``validate_dim``: Dimension of the splitting. Default are respectively $0.6$, $0.2$ and $0.2$.
#
# *Memory problem:* If ``MemoryError`` arise (with current parameters and 32GB of ram would be very unlikely), different changes can be done to make the script less RAM heavy. With `` seasonwise = True`` the regression is performed seasonally and the dataset on which the regression in performed is $1/4$ in dimension. Other matrix dimension reduction can be done maintaining the regression yearly lowering the polynomial degree of expansion (``deg``) or lowering the dimension of training dataset (``train_dim``). The latter operations reduce overall performance of the algorithm.
# In[ ]:
seasonwise = True
feature_selection = False
output_y_only = False
rnd_state=50
alphas = np.logspace(-10,5,200)
deg = 3
train_dim = 0.6
test_dim = 0.2
validate_dim = 0.2
print('-seasonwise: ',seasonwise,' -feature_selection: ',feature_selection,' -output_y_only: ',output_y_only)
seasons_list = ['spring','summer','autumn','winter']
consistency_splitting(train_dim, test_dim ,validate_dim)
model = make_pipeline(StandardScaler(),PolynomialFeatures(deg,include_bias=True), RidgeCV(alphas))
print(model)
print('')
# In[ ]:
if seasonwise==False:
feature_selection = False
# Paths of the scripts.
# In[ ]:
#Local
# DATA_FOLDER = r'../data/'
# RESULTS_FOLDER = r'../results/ridge_regression/'
#Server
DATA_FOLDER = '~/scripts/'
RESULTS_FOLDER = '/raid/motus/results/ridgeregression/'
# In[ ]:
tot_df=pd.read_csv(DATA_FOLDER+'regression_mat_year.csv',index_col=0)
# Imports of the features selected by the random forest.
# In[ ]:
if feature_selection:
df_features_sel = pd.read_csv('feature_for_ridge_season.txt',header=None)
df_features_sel = df_features_sel.drop(df_features_sel.columns[0], axis=1)
lst_features_sel=np.array(df_features_sel).tolist()
# Transform absolute value and direction in vector components
# In[ ]:
tot_df = vectorize_wind_speed(tot_df)
# Split season by season
# In[ ]:
if seasonwise:
df_spring, df_summer, df_autumn, df_winter = season_splitter(tot_df)
season_dfs=[df_spring, df_summer, df_autumn, df_winter]
else:
season_dfs=[tot_df]
# In[ ]:
#Empty dataframes
mses_results=[]
r_2s_results=[]
mag_avg_pred_results=[]
mag_avg_true_results=[]
std_true_results=[]
# In[ ]:
for index,df in enumerate(season_dfs):
if len(season_dfs)>2:
names=seasons_list
else:
names=['allyear']
#Printing progress
print('Period under optimization: ',names[index])
#Dividing X and y
X = df.drop(columns=['u_x', 'u_y','u_z'])
if feature_selection:
print('Features considered: ',lst_features_sel[index])
print('')
X=np.array(X[lst_features_sel[index]])
h_position=lst_features_sel[index].index('h')
else:
X=np.array(X)
#Chose 1D or 2D output
if output_y_only:
y = np.array(df['u_y'])
else:
y = np.array(df[['u_x', 'u_y']])
#Splitting Matrices
X_tr, X_temp, y_tr, y_temp = train_test_split(X, y, test_size=test_dim+validate_dim, random_state=rnd_state)
X_te, X_va, y_te, y_va = train_test_split(X_temp, y_temp, test_size=validate_dim/(test_dim+validate_dim), random_state=rnd_state)
X_temp = None
y_temp = None
#Make a list with differet heights
if feature_selection:
X_te_hs, y_te_hs = split_hs_test(X_te,y_te,h_pos=h_position)
else:
X_te_hs, y_te_hs = split_hs_test(X_te,y_te)
#print('X_shape: ',X.shape)
#Fit the model
model.fit(X_tr,y_tr)
y_pred_hs=[]
mag_avg_pred=[]
#Computing and saving the predictions and the average magnitude
for hs in X_te_hs:
ans=model.predict(hs)
y_pred_hs.append(ans)
ans=magnitude_avg(ans)
#print(ans.shape)
mag_avg_pred.append(ans)
mag_avg_pred.append(names[index])
mag_avg_pred_results.append(mag_avg_pred)
mses=[]
mag_avg_true=[]
std_true=[]
#Computing and saving mse, true magnitude averaged, true std
for idx,i in enumerate(y_pred_hs):
mses.append(mean_squared_error(y_te_hs[idx],i))
Ans=y_te_hs[idx]
#ans=np.sqrt(np.sum(np.square(ans),axis=1)).mean()
ans=magnitude_avg(Ans)
mag_avg_true.append(ans)
ans=std_avg(Ans)
std_true.append(ans)
mses.append(names[index])
mag_avg_true.append(names[index])
std_true.append(names[index])
mses_results.append(mses)
mag_avg_true_results.append(mag_avg_true)
std_true_results.append(std_true)
r_2s=[]
#Computing and saving r_2 metric
for idx,i in enumerate(y_pred_hs):
r_2s.append(r2_score(y_te_hs[idx],i))
r_2s.append(names[index])
r_2s_results.append(r_2s)
#Plot and save graphs. Save or show can be chosen in bolean save.
if output_y_only:
plot_ys_single(y_pred_hs,y_te_hs,RESULTS_FOLDER+'/images/',save=True,name=(names[index]+'-'))
else:
plot_ys(y_pred_hs,y_te_hs,RESULTS_FOLDER+'/images/',save=True,name=(names[index]+'-'))
# Save results in .txt in the target results folder.
# In[ ]:
mses_results_df=pd.DataFrame(mses_results)
mses_results_df.to_csv(RESULTS_FOLDER+'mses_u_seasons.txt',header=None, sep=',', mode='a')
# In[ ]:
r_2s_results_df=pd.DataFrame(r_2s_results)
r_2s_results_df.to_csv(RESULTS_FOLDER+'rsquared_u_seasons.txt',header=None, sep=',', mode='a')
# In[ ]:
pd.DataFrame(mag_avg_pred_results).to_csv(RESULTS_FOLDER+'magnitude_average_pred.txt',header=None, sep=',', mode='a')
pd.DataFrame(mag_avg_true_results).to_csv(RESULTS_FOLDER+'magnitude_average_true.txt',header=None, sep=',', mode='a')
pd.DataFrame(std_true_results).to_csv(RESULTS_FOLDER+'magnitude_std_true.txt',header=None, sep=',', mode='a')
# In[ ]:
print('JOB Finished')
# Plot names and title on the picture itself explain their content. An interval is chosen randomly to visualize the behaviour on the true value and the prediction. The MSE, $R^2$, average prediction, average true values and average standard deviations are all saved on ``.txt`` format. The last entry of each line identifies the period (season or all year), while every number not considering the first one is referring to the mast anemometer from 1 to 6 in this order.

Event Timeline

c4science · Help